AKIO HIRAYAMA · Field guide

The Handoff & Approval Card: a field guide

How to decide when AI work is actually finished, and how to pass it from one AI to another without relying on trust.

The card itself lives in the multi-ai-handoff-card repository. This page explains how to use it: what each field is for, what a filled-in card looks like, and where it tends to go wrong.

Why a card at all

An AI saying "done" is a claim. A card turns that claim into something another person, script or AI can check. The rule is simple: if a field is empty, the work is not done. It is only reported.

Field by field

Field	What to write	Common mistake
Task	The request in one sentence, as it was asked	Rewriting the task to match what was done
Claimed result	What the AI says it did, quoted or summarised	Mixing the claim with the evidence
Risk tier	auto / auto-then-verify / human approval, and why	Choosing a tier after the work is already live
Evidence	What was checked, by what, and the result. Not the same AI run	"The AI confirmed it" counted as evidence
Changed	Pages, files or systems touched	Listing only the intended changes
Rollback	How to undo it, and who can	"Revert if needed" with no method
Open items	Anything unverified, ambiguous or skipped	Leaving it blank to look finished
Approved by	A named human, or the rule that allowed auto-approval	Approval implied by silence

A worked example (hypothetical)

This example is invented to show the format. It does not describe a real incident.

TASK:            Translate the pricing FAQ into Spanish on a draft page
CLAIMED RESULT:  "All 12 answers translated and saved as a draft"
RISK TIER:       auto-then-verify (draft only; nothing public changes)
EVIDENCE:        Script counted 12 answers in source and 12 in draft: pass
                 Second model, fresh context, compared numbers and
                 currencies: 1 mismatch (EUR shown as USD): fail
CHANGED:         Draft page only
ROLLBACK:        Delete the draft; the live page was not touched
OPEN ITEMS:      Fix the currency line; native review of tone not done
APPROVED BY:     Not yet. Needs a human before publishing

Note what the card does here: the AI's claim was mostly true, but the evidence field shows one failed check and the open items keep it from being called finished.

Machine-readable version

The same card as JSON, for pipelines and for handing work between AIs:

{
  "task": "...",
  "claimed_result": "...",
  "risk_tier": "auto | auto-then-verify | human-approval",
  "risk_reason": "...",
  "evidence": [{"check": "...", "checked_by": "...", "result": "pass | fail | unclear"}],
  "changed": ["..."],
  "rollback": {"how": "...", "who": "..."},
  "open_items": ["..."],
  "approved_by": "..."
}

Download it from the repository: card.json.

Where it still falls short

A passing check is evidence, not proof. Verifiers can be wrong too.
A different AI is not automatically an independent one.
Risk tiers drift when the context of an action changes.

More from AKIO HIRAYAMA

Official website: akiohirayama.com · Notes: Multi-AI Control Notes · Medium · GitHub
AKIO HIRAYAMA is hiring an AI Engineer (USD 30,000 per month) to build this kind of control layer: role details